Early detection of downy mildew in vineyards using deep neural networks for semantic segmentation
Inés Hernández, Rui Silva, Pedro Melo‐Pinto, Salvador Gutiérrez, Javier Tardáguila
Abstract
Downy mildew is a critical disease in viticulture, typically identified through manual inspection of individual leaves in the field by experts. The combination of artificial intelligence techniques with mobile platforms can optimise non-invasive detection. This work focused on employing semantic segmentation deep neural networks to detect visual symptoms of downy mildew in high-resolution grapevine images under field conditions. Vineyard canopy images were collected from 14 plots using both manual and mobile platform methods. The study compared six architectures and six encoders using transfer learning, as well as two SegNet AdHoc architectures. To address imbalance problems, simple data augmentation, MixUp, oversampling, and undersampling techniques were employed. The results were adjusted through test-time augmentation. The study found that the U-Net architecture, using the MobileVit-S encoder and the Dice loss function, was particularly efficient. The U-Net architecture with light-weight encoders exhibited potential for real-time applications. The robustness of the model was improved by combining oversampling and undersampling with simple data augmentation during training. The classification of areas with and without disease symptoms achieved an accuracy of 86% and an f1-score of 82%. Additionally, the number of symptoms in grapevine canopy images was detected with an NRMSE of 12%. In conclusion, the proposed methodology shows promise for efficiently early assessing grapevine downy mildew under field conditions. This approach could be applied to other crop diseases and pests, taking advantage of the complexity of the dataset to strengthen the robustness of the model in real-world scenarios. • Non-invasive early detection of downy mildew in vineyards. • State-of-the-art semantic segmentation networks used to localise disease symptoms. • Real-time potential using small neural networks. • Increasing of model robustness by data augmentation techniques. • Reduction of false positive values with Test-Time-Augmentation.